Application Of Multivariable Statistical Method On Fault Diagnosis

In modern process industries, with the rapid development of mass production and complexity, reliability and security are being greatly needed to avoid large economical loss brought by accidents and even breakdowns of industrial productions. Therefore, it is very important for industrial integrated automation to research and develop process supervisory systems which integrate the functions of control, supervision and diagnosis. Fault detection (FD) technology supervises the status of production process, detect the fluctuation and faults, and locate the fault sources, then isolate and eliminate them. In this way, calamities are prevented and fluctuations of quality of products are reduced.In this thesis, important aspects of process fault detection based on multivariate statistical theory, which is an important branch of FD technology are presented and studied systematically. The main work could be stated as follows:In Chapter 1, the main contents of FD, the classification of various detection methods and prospects are summarized, especially the history and status about detection methods based on statistical theory.In Chapter 2, the basis of statistical detection methods are introduced: Principal Component Analysis and Partial Least Squares, and their extensions: Multi-way Principal Component Analysis, Multi-block Partial Least Squares.In Chapter 3, an improved PCA is presented which calculates the principal subspace of data collected from current time window and compares it with that from data collected under a normal condition. Both the improved PCA and conventional PCA are used to monitor the Tennessee Eastman process, the improved PCA has better performance to detect weak process changes.In Chapter 4, fault isolation based on statistical methods is discussed. Firstly, the fault isolation ability of Multi-block Principal Component Analysis istested on the Tennessee Eastman process. Then Principal Component Analysis and Signed Directed Graph are combined, and the method is applied to a Continuous Stirred Tank Reactor.Chapter 5 concludes with a summary and discussions of future and prospective research on open problems.